How to Get Good Performance by Using OpenMP ! ! Loop optimizations ! - - PowerPoint PPT Presentation
Agenda ! How to Get Good Performance by Using OpenMP ! ! Loop optimizations ! ! Measuring OpenMP performance ! ! Best practices ! ! Task parallelism !! !" #" Memory access patterns ! Correctness versus performance ! It
!"
#" $"
%"
&"
'"
("
)"
*"
!+"
cont'd on next page!
n
!!" !#"
3.2
!Returns the wall clock time (in seconds) relative to ! !an arbitrary reference time. !
!$"
The amount of time required to coordinate parallel threads, as
Parallel overhead can include factors such as:!
Thread start-up time!
Synchronization!
Software overhead imposed by parallel compilers, !libraries, tools, operating system, etc.!
Thread termination time!
!%"
Speedup(P) = TSerial(P) TElapsed(P) = 1 f P ! f +1+ OP " P = 1 0.95 P + 0.05 + 0.02 " P Efficiency(P) = Speedup(P) P
!&" !'"
!handling OpenMP constructs.!
!synchronization points.!
!critical regions.!
!("
,-./01234" 56,677-7"83," 56,677-7" 924:7-" ;6,,2-," 83," !)"
!*"
#+"
#!"
##"
#$"
#%"
If at all possible, an atomic update is to be preferred.!
#&"
!"#
#pragma omp parallel { #pragma omp critical { ... } ... }
time!
#'"
#("
Large parallel regions offer more opportunities for using data in cache and provide a bigger context for compiler optimizations.!
#)"
#*"
#*"
time! #pragma omp parallel { #pragma omp for for ( ; ; ) { } }
$+"
!balancing is not an issue!
!distributing the work over the threads!
! ! ! ! ! ! !- multiplication of triangular matrices ! ! ! ! ! !- parallel searches in a linked list!
!various iteration scheduling algorithms !
$!"
cont'd on next page!
$#"
$$"
The state bit of a cache line does not keep track of the cache line state on a byte basis, but at the line level instead.! Thus, if independent data items happen to reside on the same cache line (cache block), each update will cause the cache line to “ping-pong” between the threads. ! ! ! ! ! ! ! ! ! ! ! ! ! ! This is called false sharing.!
$%"
False sharing is likely to significantly impact performance under the following conditions:!
the same cache line(s).!
$&"
$'"
Array elements are contiguous in memory and hence share cache lines.! ! !! ! !Result: False sharing may lead to poor scaling! Solutions:!
!with local copies of the array in stead of an array !indexed by the thread ID.!
!cache lines.!
$(" $("
int a[Nthreads]; #pragma omp parallel for shared(Nthreads,a) schedule(static,1) for (int i=0; i<Nthreads; i++) a[i] += i; int a[Nthreads][cache_line_size]; #pragma omp parallel for shared(Nthreads,a) schedule(static,1) for (int i=0; i<Nthreads; i++) a[i][0] += i;
$)"
$*" %+"
%!" %#" %$" %%"
$"#
Tasks are independent units of work! Threads are assigned to perform the work of each task!
The runtime system decides which of the above!
%'"
OpenMP has always had tasks, but they were not called that.!
!implicit tasks, one per thread.!
OpenMP 3.0 adds a way to create a task explicitly for the team to execute.!
%("
#pragma omp task [clause [[,] clause] ... ] structured block
Each encountering thread creates a new task.!
An OpenMP barrier (implicit or explicit):! ! ! !All tasks created by any thread of the current !team are guaranteed to be completed at barrier exit.! Task barrier (taskwait): ! ! ! ! !Encountering thread suspends until all child tasks it !has generated are !complete.!
%)"
void process_list(elem_t *elem) { #pragma omp parallel { #pragma omp single { while (elem != NULL) { #pragma omp task { process(elem); } elem = elem->next; } } } } elem is firstprivate by default!
%*"
int fib(int n) { int i, j; if (n < 2) return n; i = fib(n - 1); j = fib(n - 2); return i + j; } int main() { int n = 10; printf("fib(%d) = %d\n", n, fib(n)); } #pragma omp parallel #pragma omp single #pragma omp task shared(i) #pragma omp task shared(j) #pragma omp taskwait
Computation of Fibonacci numbers! 1,1,2,3,5,8,13,21,34,55,89,144,...!
&+"
struct node { struct node *left, *right; }; void traverse(struct node *p, int postorder) { if (p->left != NULL) #pragma omp task traverse(p->left, postorder); if (p->right != NULL) #pragma omp task traverse(p->right, postorder); if (postorder) { #pragma omp taskwait } process(p); }
&!"
Certain constructs have suspend/resume points at defined positions within them.! When a thread encounters a suspend point, it is allowed to suspend the current task and resume another. It can then return to the original task and resume it.! A tied task must be resumed by the same task that suspended it.! Tasks are tied by default. A task can be specified to be untied using! ! ! !#pragma omp task untied
&#"
The collapse clause (in OpenMP 3.0) handles perfectly nested multi-dimensional loops.!
#pragma omp for collapse(2) for (i = 0; i < N; i++) for (j = 0; j < M; j++) for (k = 0; k < K; k++) foo(i, j, k);
Iteration space from i-loop and j-loop is collapsed into a single
iteration space. !
&$"
for (i = 0; i < n; i++) { x = (b[i] + c[i]) / 2; a[i] = a[i + 1] + x; }
Serial version containing anti dependency!
#pragma omp parallel for shared(a,a_copy) for (i = 0; i < n; i++) a_copy[i] = a[i + 1]; #pragma omp parallel for shared(a,a_copy) private(x) for (i = 0; i < n; i++) { x = (b[i] + c[i]) / 2; a[i] = a_copy[i] + x; }
Parallel version with dependencies removed!
&%"
for (i = 1; i < n; i++) { b[i] = b[i] + a[i - 1]; a[i] = a[i] + c[i]; }
Serial version containing flow dependency!
b[1] = b[1] - a[0]; #pragma omp parallel for shared(a,b,c) for (i = 1; i < n; i++) { a[i] = a[i] + c[i];
} a[n - 1] = a[n - 1] + c[n - 1];
Parallel version with dependencies removed by loop skewing!
&&"
Some compilers can insert OpenMP can optimize a program
programs spread over several files create difficulties.! The Intel compilers support automatic parallelization. Example,!
icc -o matmult -O3 -parallel -par-report3 matmult.c